Juggler

Product Introduction

1. Definition: Juggler is a native desktop application and headless server, built in Go, that functions as a visual integrated development environment (IDE) and workbench for AI-powered coding agents. It is a persistent, multi-client session manager for human-AI collaboration.
2. Core Value Proposition: Juggler exists to provide developers with granular visibility and control over AI coding sessions. It replaces opaque, linear chat transcripts with a structured, inspectable, and branchable conversation tree, enabling more complex and manageable AI-assisted programming workflows locally or on remote development machines.

Main Features

1. Miller-Column Navigation & Persistent Conversation Trees: The interface uses a Finder-style Miller column layout to visualize the hierarchical structure of an AI coding session. Unlike scrolling logs, conversations are persistent trees stored on disk as Yjs documents. Users can create sub-threads at any point for delegated tasks or experiments, keeping intermediate work out of the main context and returning only final results.
2. Inspectable Model Transactions & Tool Calls: Every interaction with the Large Language Model (LLM) is fully inspectable. Users can open any model transaction to audit the exact system prompt, message history, available tool schemas, raw output, token usage, latency, and stop reason. Tool calls (e.g., file read/write, bash) open in dedicated views showing arguments, approvals, and results.
3. Multi-Client Sync & Remote Session Hosting: The core is a Go server that owns the session state. The native desktop app bundles this server for local use. For remote development, the headless juggler binary can be run on a dev box or server. Any number of clients (desktop app or web browser) can connect to the same live session, with all views synchronized in real-time via Yjs CRDTs.
4. Extensible JavaScript Plugin Architecture: All core functionality is built on a public JavaScript Extension SDK. Context items, LLM loop strategies, slash commands, file viewers, and tools are implemented as extensions. This allows users to inspect, modify, or create new plugins. The system also features fully inspectable integration for Model Context Protocol (MCP) servers.
5. Context-Aware Window Management & Pre-emptive Truncation: Juggler actively manages the LLM's context window. It calculates the total token size of the assembled request before each API call, reserves space for the response, and automatically compacts or summarizes older history when a conversation approaches the model's limit, preventing unexpected truncation errors.

Problems Solved

1. Pain Point: Lack of transparency and control in AI coding assistants. Traditional AI coding tools operate as "black boxes" with linear, unscalable chat histories, making it difficult to debug agent reasoning, manage long conversations, or delegate sub-tasks without polluting the main context.
2. Target Audience: Software engineers, DevOps professionals, and technical leads who regularly use AI coding agents (like GitHub Copilot, Claude Code, or Cursor) for complex, multi-step development tasks, refactoring, or remote development work.
3. Use Cases: Debugging a failing CI/CD script by creating a sub-thread for the agent to analyze logs separately. Refactoring a large codebase while using branches to test different architectural approaches. Working on a remote development server via a browser while keeping the session alive and persistent. Auditing an agent's work by inspecting every tool call and LLM prompt for security or quality assurance.

Unique Advantages

1. Differentiation: Unlike browser-based chat interfaces or IDE plugins, Juggler is a dedicated, persistent workbench. It contrasts with solutions like Cursor or Windsurf by offering deep transaction inspection, true conversation branching, and a client-server model for remote work, all within a native, non-Electron application.
2. Key Innovation: The combination of a Yjs-based persistent tree document model for conversations with a fully inspectable, extension-first architecture. This allows the session state to be a first-class, manipulatable data structure, while the open SDK ensures every component—from the UI to the tools the AI uses—can be audited and customized.

Frequently Asked Questions (FAQ)

1. Is Juggler an AI model or a chatbot? No, Juggler is not an AI model. It is a visual workbench and client for existing AI coding agents and LLMs like Claude, GPT-4, Gemini, and Ollama. You must bring your own API keys or subscriptions.
2. How does Juggler handle privacy and data security with my code? Juggler runs the LLM API calls and tool operations from its server (local or remote). Your code, prompts, and conversation data never pass through a third-party analytics service. The session data is stored locally on the machine running the Juggler server.
3. Can I use Juggler with GitHub Copilot? Yes, Juggler supports GitHub Copilot as a provider via its API, allowing you to use your Copilot subscription within Juggler's structured, inspectable interface alongside other models.
4. What does "open-source core" mean for extensions? The main Juggler application is licensed under AGPLv3. However, the Extension SDK and bundled example extensions are licensed under the permissive Apache 2.0 license, meaning you can write proprietary, closed-source extensions for Juggler.
5. Do I need to use the terminal to run Juggler? For typical local use, no. The desktop application for macOS, Windows, and Linux is a native binary that launches its integrated server automatically. The terminal is only needed to run the headless server binary for remote or custom deployment scenarios.